Learning from Human Teachers: Issues and Challenges for ILP in Bootstrap Learning Sriraam Natarajan 1, Gautam Kunapuli 1, Richard Maclin 3, David Page 1, Ciaran O’Reilly 2, Trevor Walker 1 and Jude Shavlik 1 2 SRI International 1 University of Wisconsin-Madison 3 University of Minnesota, Duluth

Outline Bootstrap Learning (BL), Inductive Logic Programming (ILP) and Challenges The Wisconsin Relevance Interpreter: Incorporating teacher advice into an ILP agent The Onion: Automating agent’s ILP runs across tasks and domains without expert intervention Conclusions

Outline Bootstrap Learning (BL), Inductive Logic Programming (ILP) and Challenges The Wisconsin Relevance Interpreter: Incorporating teacher advice into an ILP agent The Onion: Automating agent’s ILP runs across tasks and domains without expert intervention Conclusions

Bootstrap Learning New learning paradigm by Oblinger (2006) – proposes framework to create an electronic student that learns from natural instruction – views learning as knowledge acquisition from a human teacher Image Source: Dan Oblinger, AAAI 2006 Briefing

Bootstrap Learning electronic student (agent) assumes relevant knowledge possessed by teacher – domain background teaching through human-like natural instruction methods – pedagogical examples – telling of instructions – demonstration – feedback

Bootstrap Learning Learning depends only on instruction method, is domain- independent  Unmanned Aerial Vehicle (UAV)  Armored Task Force (ATF)  International Space Station (ISS) Student learns concepts through a ladder of lessons more complex concepts bootstrap on lower-rung lessons simpler concepts; learned first lesson ladder

Bootstrap Learning From Examples We focus on instruction by examples – including teacher hints about specific examples – teacher and student communicate via Interlingua Learn models using Inductive Logic Programming (ILP, Muggleton and de Raedt, 1994) – declarative representation of examples and learned rules makes it easier for teacher and student to communicate – concepts learned earlier can be bootstrapped easily We have developed a Java-based ILP system called WILL: the Wisconsin Inductive Logic Learner

Bootstrap Learning From Examples We focus on instruction by examples – including teacher hints about specific examples – teacher and student communicate via Interlingua Learn models using Inductive Logic Programming (ILP, Muggleton and De Raedt, 1994) – declarative representation of examples and learned rules makes it easier for teacher and student to communicate – concepts learned earlier can be bootstrapped easily We have developed a Java-based ILP system called WILL: the Wisconsin Inductive Logic Learner

Bootstrap Learning From Examples We focus on instruction by examples – including teacher hints about specific examples – teacher and student communicate via Interlingua Learn models using Inductive Logic Programming (ILP, Muggleton and De Raedt, 1994) – declarative representation of examples and learned rules makes it easier for teacher and student to communicate – concepts learned earlier can be bootstrapped easily We have developed a Java-based ILP system called WILL: the Wisconsin Inductive Logic Learner

Recognize “interesting scenario” tasks UAV Domain with ByExample Lessons FullFuelTankAssessGoal Near TruckAtIntersection ReadyToFly IsStopped TruckScenario IsMoving TruckScenario ComputeStoppedTrucksIntComputeMovingTrucksInt ComputeScenarioInterestingness Fly and land UAV tasks Surveillance and Camera Tasks Goal: Train an Unmanned Aerial Vehicle agent to perform Intelligent Surveillance and Reconnaissance (ISR)

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Teacher utters information about concept being taught and the natural instruction method

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Perceptions provide information about the current state

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) The teacher can gesture at perceived objects. They will be designated “this” by default.

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Teacher gives label information, identifying this as a NEGATIVE example

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample)) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Teacher gives additional “hints” about why this is a positive or negative example. This is the relevance information.

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Some more examples follow

Natural Instruction of the Agent lower rung task: FullFuelTank (determine if fuel tank is full) u: LessonTeaches(WhenTrue(Full)) u: Method(byExample) p: UAV(name=uav1, altitude=4.0, fuelCapacity=500, currentFuel=400) u: GestureAt(uav1) u: Not(Full(this)) u: Relevant(Of(fuelCapacity, this)) u: Relevant(Of(currentFuel, this))... p: UAV(name=uav2, altitude=0.0, fuelCapacity=300, currentFuel=300) i: Answer(response=true) i: Grade(grade=100.0) Teacher tests the student and gives a grade

Representation for ILP An ILP system learns a logic program given – background knowledge as a set of first-order logic (FOL) formulae – examples expressed as facts represented in logic terms are objects in the world – constants: uav1, uav2 – variables: ?uav – functions: CurrentFuel(?uav, ?currentFuel) literals are truth-valued and represent relations among objects –sameAs(?fuelCapacity, ?currentFuel) literals can be combined into compound sentences using logical connectives AND, OR and NOT. We translate from Interlingua to an ILP representation creating background, facts and examples from messages

Representation for ILP An ILP system learns a logic program given – background knowledge as a set of first-order logic (FOL) formulae – examples expressed as facts represented in logic terms are objects in the world – constants: uav1, uav2 – variables: ?uav – functions: CurrentFuel(?uav, ?currentFuel) literals are truth-valued and represent relations among objects –sameAs(?fuelCapacity, ?currentFuel) literals can be combined into compound sentences using logical connectives AND, OR and NOT. We translate from Interlingua to an ILP representation creating background, facts and examples from messages

Representation for ILP An ILP system learns a logic program given – background knowledge as a set of first-order logic (FOL) formulae – examples expressed as facts represented in logic terms are objects in the world – constants: uav1, uav2 – variables: ?uav – functions: CurrentFuel(?uav, ?currentFuel) literals are truth-valued and represent relations among objects –sameAs(?fuelCapacity, ?currentFuel) literals can be combined into compound sentences using logical connectives AND, OR and NOT. We translate from Interlingua to an ILP representation creating background, facts and examples from messages

Representation for ILP An ILP system learns a logic program given – background knowledge as a set of first-order logic (FOL) formulae – examples expressed as facts represented in logic terms are objects in the world – constants: uav1, uav2 – variables: ?uav – functions: CurrentFuel(?uav, ?currentFuel) literals are truth-valued and represent relations among objects –sameAs(?fuelCapacity, ?currentFuel) literals can be combined into compound sentences using logical connectives AND, OR and NOT. We translate from Interlingua to an ILP representation creating background, facts and examples from messages

Representation for ILP An ILP system learns a logic program given – background knowledge as a set of first-order logic (FOL) formulae – examples expressed as facts represented in logic terms are objects in the world – constants: uav1, uav2 – variables: ?uav – functions: CurrentFuel(?uav, ?currentFuel) literals are truth-valued and represent relations among objects –sameAs(?fuelCapacity, ?currentFuel) literals can be combined into compound sentences using logical connectives AND, OR and NOT. We translate from Interlingua to an ILP representation creating background, facts and examples from messages

Learning in ILP Features: A..Z Target Concept to be learned: A,Z,R,W / Target Have background, examples and facts for learning ILP search adds one literal at a time n predicates leads to O(n!) combos ILP requires large number of examples Hypothesis space True / Target A,Z / TargetA,B / Target A / Target A, Z, R, W / Target A, Z, R / Target...

Learning in ILP with Relevance Teacher tells agent that p redicates A, Z, R are relevant to the concept being learned amount of search is greatly reduced ILP with relevance can succeed with fewer examples Hypothesis space True / Target A,Z / TargetA,B / Target A / Target A, Z, R, W / Target A, Z, R / Target... Features: A..Z Target Concept to be learned: A,Z,R,W / Target

BL/ILP Challenges Relevant statements specify partial theory – how can we use relevance effectively in ILP? No ILP expert in the loop for param selection – agent should be trainable by non-ILP experts – agent should be able to use ILP over different lessons in the same domain AND across different domains – cannot exhaustively try all possible combinations Teacher is agnostic towards learning mechanism – examples may contain a very large number of features – #examples very low (< 10 in most lessons) – some lessons have no negative examples

BL/ILP Challenges Relevant statements specify partial theory – how can we use relevance effectively in ILP? No ILP expert in the loop for param selection – agent should be trainable by non-ILP experts – agent should be able to use ILP over different lessons in the same domain AND across different domains – cannot exhaustively try all possible combinations Teacher is agnostic towards learning mechanism – examples may contain a very large number of features – #examples very low (< 10 in most lessons) – some lessons have no negative examples

BL/ILP Challenges Relevant statements specify partial theory – how can we use relevance effectively in ILP? No ILP expert in the loop for param selection – agent should be trainable by non-ILP experts – agent should be able to use ILP over different lessons in the same domain AND across different domains – cannot exhaustively try all possible combinations Teacher is agnostic towards learning mechanism – examples may contain a very large number of features – #examples very low (< 10 in most lessons) – some lessons have no negative examples

BL/ILP Challenges Relevant statements specify partial theory – Develop a robust strategy to interpret and incorporate relevance ( Wisconsin Relevance Interpreter) No ILP expert in the loop for param selection – Develop a multi-layer strategy for automating ILP runs so agent can be easily trained, generalized across domains ( Onion) Teacher is agnostic towards learning mechanism – Develop novel approaches to handle non-standard cases e.g., generation of negative examples etc.

BL/ILP Challenges Relevant statements specify partial theory – Develop a robust strategy to interpret and incorporate relevance ( Wisconsin Relevance Interpreter) No ILP expert in the loop for param selection – Develop a multi-layer strategy for automating ILP runs so agent can be easily trained, generalized across domains ( Onion) Teacher is agnostic towards learning mechanism – Develop novel approaches to handle non-standard cases e.g., generation of negative examples etc.

BL/ILP Challenges Relevant statements specify partial theory – Develop a robust strategy to interpret and incorporate relevance ( Wisconsin Relevance Interpreter) No ILP expert in the loop for param selection – Develop a multi-layer strategy for automating ILP runs so agent can be easily trained, generalized across domains ( Onion) Teacher is agnostic towards learning mechanism – Develop novel approaches to handle non-standard cases e.g., generation of negative examples etc.

Outline Bootstrap Learning (BL), Inductive Logic Programming (ILP) and Challenges The Wisconsin Relevance Interpreter: Incorporating teacher advice into an ILP agent The Onion: Automating agent’s ILP runs across tasks and domains without expert intervention Conclusions

Handling Teacher Hints relevance can be specified at different “resolutions” – certain constants are relevant (e.g.,  ) – attribute (incl. type) is relevant (e. g., UAV.currentFuel) – object is relevant (e.g., Truck3) – relationship among attributes/objects is relevant (e.g., lessThan, greaterThan, sameAs) – previously learned concepts (bootstrapping) – partial theory (e.g., combinations of the above) teacher gives hints about specific examples – key challenge: what should be generalized (to a logical variable) and what should remain constant?

Handling Teacher Hints relevance can be specified at different “resolutions” – certain constants are relevant (e.g.,  ) – attribute (incl. type) is relevant (e. g., UAV.currentFuel) – object is relevant (e.g., Truck3) – relationship among attributes/objects is relevant (e.g., lessThan, greaterThan, sameAs) – previously learned concepts (bootstrapping) – partial theory (e.g., combinations of the above) teacher gives hints about specific examples – key challenge: what should be generalized (to a logical variable) and what should remain constant?

Handling Teacher Hints relevance can be specified at different “resolutions” – certain constants are relevant (e.g.,  ) – attribute (incl. type) is relevant (e. g., UAV.currentFuel) – object is relevant (e.g., Truck3) – relationship among attributes/objects is relevant (e.g., lessThan, greaterThan, sameAs) – previously learned concepts (bootstrapping) – partial theory (e.g., combinations of the above) teacher gives hints about specific examples – key challenge: what should be generalized (to a logical variable) and what should remain constant?

Handling Teacher Hints relevance can be specified at different “resolutions” – certain constants are relevant (e.g.,  ) – attribute (incl. type) is relevant (e. g., UAV.currentFuel) – object is relevant (e.g., Truck3) – relationship among attributes/objects is relevant (e.g., lessThan, greaterThan, sameAs) – previously learned concepts (bootstrapping) – partial theory (e.g., combinations of the above) teacher gives hints about specific examples – key challenge: what should be generalized (to a logical variable) and what should remain constant?

Handling Teacher Hints relevance can be specified at different “resolutions” – certain constants are relevant (e.g.,  ) – attribute (incl. type) is relevant (e. g., UAV.currentFuel) – object is relevant (e.g., Truck3) – relationship among attributes/objects is relevant (e.g., lessThan, greaterThan, sameAs) – previously learned concepts (bootstrapping) – partial theory (e.g., combinations of the above) teacher gives hints about specific examples – key challenge: what should be generalized (to a logical variable) and what should remain constant?

Handling Relevance: One Example identify “interesting” candidate constants isInteresting(Truck17) isInteresting(Stopped) identify “relevant” predicates for the target concept relevant: Truck.moveStatus relevant: Truck relevant: sameAs (these get lower cost during search) GestureAt(Truck(name=Truck17, latitude=-10, longitude=10, moveStatus=“Stopped”)) RelevantRelationship(InstanceOf(this, Truck)) RelevantRelationship(SameAs(this.moveStatus, “Stopped”))

construct rules for each relevance statement pred1(?T) IF Truck(?T) pred2(?T) IF moveStatus(?T, ?S), sameAs(?S, stopped) create combo rule from all the relevance for this example relevantFromPosEx1(?T) IF pred1(?T), pred2(?T) Handling Relevance: One Example GestureAt(Truck(name=Truck17, latitude=-10, longitude=10, moveStatus=“Stopped”)) RelevantRelationship(InstanceOf(this, Truck)) RelevantRelationship(SameAs(this.moveStatus, “Stopped”))

construct combinations of rules from all pos examples posCombo(?T) IF relevantFromPosEx1(?T), relevantFromPosEx2(?T),..., relevantFromPosExN(?T) and neg examples negCombo(?T) IF ~relevantFromNegEx1(?T), ~relevantFromNegEx2(?T),..., ~relevantFromNegExN(?T) GestureAt(Car12) // Negative Example relevance RelevantRelationship(arg1 = Not(InstanceOf(arg1 = this, arg2 = Truck))) Handling Relevance: Across All Examples

construct all comboRules: cross products across different combos allCombo(?T) IF posCombo(?T), negCombo(?T) all predicates (relevantFrom, posCombo, negCombo, allCombo) added to background during search also combine advice about examples using OR Handling Relevance: Across All Examples We use both AND and OR because a human teacher might teach – a conjunctive concept with each example illustrating only one piece of the concept, or – a disjunctive concept with several alternatives, with each alternative illustrated via a different example. – Added bonus: robustness to teacher errors

construct all comboRules: cross products across different combos allCombo(?T) IF posCombo(?T), negCombo(?T) all predicates (relevantFrom, posCombo, negCombo, allCombo) added to background during search also combine advice about examples using OR Handling Relevance: Across All Examples We use both AND and OR because a human teacher might teach – a conjunctive concept with each example illustrating only one piece of the concept, or – a disjunctive concept with several alternatives, with each alternative illustrated via a different example. – Added bonus: robustness to teacher errors

Outline Bootstrap Learning (BL), Inductive Logic Programming (ILP) and Challenges The Wisconsin Relevance Interpreter: Incorporating teacher advice into an ILP agent The Onion: Automating agent’s ILP runs across tasks and domains without expert intervention Conclusions

The Need for Automation supervised learning methods require – intervention by domain experts for feature selection, predicate design – intervention by machine learning experts for parameter selection, guiding search automation of ILP setup is required for – usability by non-expert human teachers – task and domain independence – standard automation methods (such as CV) not easily applicable because of small number of examples

The Need for Automation supervised learning methods require – intervention by domain experts for feature selection, predicate design – intervention by machine learning experts for parameter selection, guiding search automation of ILP setup is required for – use by non-expert human teachers – task and domain independence – standard automation methods (such as CV) not easily applicable because of small number of examples

ONION: A Multi-Layer Strategy Automating the ILP Setup Problem start with a small search space and incrementally expand until a “good theory” is found during the search – prefer simple concepts over complex concepts – prefer highly relevant sub-concepts (exploit relevance) – try various parameter settings for best theory a “good theory” covers most of positive examples and very few negative examples level of coverage can be adjusted to find a theory of acceptable accuracy

ONION: A Multi-Layer Strategy Automating the ILP Setup Problem start with a small search space and incrementally expand until a “good theory” is found during the search – prefer simple concepts over complex concepts – prefer highly relevant sub-concepts (exploit relevance) – try various parameter settings for best theory a “good theory” covers most of positive examples and very few negative examples level of coverage can be adjusted to find a theory of acceptable accuracy

ONION: A Multi-Layer Strategy Automating the ILP Setup Problem start with a small search space and incrementally expand until a “good theory” is found during the search – prefer simple concepts over complex concepts – prefer highly relevant sub-concepts (exploit relevance) – try various parameter settings for best theory a “good theory” covers most of positive examples and very few negative examples level of coverage can be adjusted to find a theory of acceptable accuracy

ONION: A Multi-Layer Strategy short rules; highly relevant combo-rules medium-length rules; all relevant features flip example labels (learn negation of concept) consider small subset of parameter settings long rules; all features More details in longer technical paper (Walker et al., 2010)

Outline Bootstrap Learning (BL), Inductive Logic Programming (ILP) and Challenges The Wisconsin Relevance Interpreter: Incorporating teacher advice into an ILP agent The Onion: Automating agent’s ILP runs across tasks and domains without expert intervention Results and Conclusions

Experimental Results Domains and data for agent – 3 domains, 56 lessons (concepts) – 7.6 labeled examples per concept on average – mixture of positives and negatives – teacher gives relevant information for each concept Results – without relevance: average accuracy = 63.9% – with relevance: average accuracy = 100% – no expert intervention – further robustness experiments currently in progress

Experimental Results Domains and data for agent – 3 domains, 56 lessons (concepts) – 7.6 labeled examples per concept on average – mixture of positives and negatives – teacher gives relevant information for each concept Results – without relevance: average test-set accuracy = 63.9% – with relevance: average test-set accuracy = 100% – “mystery domain”: average test-set accuracy = 100% – further robustness experiments currently in progress

Conclusions Bootstrap Learning agent interacts with teacher using natural instruction methods – gets examples, background, relevance information Teacher provided relevance can dramatically reduce – ILP search needed – number of examples needed Challenge #1: Generalizing advice about specific examples (Solution: Wisconsin Relevance Interpreter) Challenge #2: Automating ILP runs for wider usability, and task- and domain-independence (Solution: Onion)

Conclusions Bootstrap Learning agent interacts with teacher using natural instruction methods – gets examples, background, relevance information Teacher provided relevance can dramatically reduce – ILP search needed – number of examples needed Challenge #1: Generalizing advice about specific examples (Solution: Wisconsin Relevance Interpreter) Challenge #2: Automating ILP runs for wider usability, and task- and domain-independence (Solution: Onion)

Questions? Acknowledgements: The authors gratefully acknowledge support of the Defense Advanced Research Projects Agency under DARPA grant FA C Views and conclusions contained in this document are those of the authors and do not necessarily represent the official opinion or policies, either expressed or implied of the US government or of DARPA. References: (Oblinger, 2006) Oblinger, D. Bootstrap learning - external materials (Muggleton and De Raedt, 1994) Muggleton, S. and De Raedt, L. Inductive logic programming: Theory and methods. Journal of Logic Programming, 19/20:629–679, (Walker et al., 2010) Walker, T., Kunapuli, G., Natarajan, S., Shavlik, J. W. and Page, D. Automating the ILP Setup Task: Converting User Advice about Specific Examples into General Background Knowledge. International Conference on Inductive Logic Programming, (to appear)